MedGemma 1.5 Open‑Sourced

Google DeepMind has released MedGemma 1.5, an open medical artificial intelligence model that can read both clinical text and medical images. (youtube.com) Medical language models work like autocomplete systems trained on health data, while image encoders turn scans or photos into numbers a model can compare and reason over. DeepMind said MedGemma 1.5 combines the Gemma 3 model family with a MedSigLIP image encoder so one system can handle both kinds of input. (youtube.com) The release was presented in a demonstration video that showed the model working across imaging, notes and other healthcare data streams instead of one narrow benchmark. DeepMind framed that setup as a fit for clinical workflows, where doctors move between pictures, reports and written history. (youtube.com) Google has been pushing smaller open models through the Gemma line since February 2024, when it introduced Gemma as a family built from the same research behind Gemini. In March 2025, Google expanded that line with Gemma 3, which it described as a multimodal model family for text and vision tasks. (blog.google ) (blog.google) That matters in medicine because hospitals do not store information in one format. A chest X-ray, a pathology slide, a discharge summary and a referral note often sit in separate systems, and a model that can process both words and images is designed for that split. (youtube.com) Open release also changes who can test the model. Instead of using only a closed application programming interface, researchers and health systems can download weights, fine-tune the system on local data and build tools around their own compliance and security rules. (huggingface.co) Google has been making a similar argument with other health models. In May 2024, DeepMind and Google Research published Med-Gemini, a medical system aimed at text, images and long clinical context, but that work centered on a research paper rather than an open-weight release. (research.google) The company is not alone in that race. Microsoft, Meta and a range of academic groups have been building medical vision-language models, but deployment still runs into the same limits: privacy law, uneven hospital data and the need to prove performance on real patients rather than demo cases. (nature.com) (fda.gov) United States regulators have been drawing that line for years. The Food and Drug Administration says artificial intelligence tools used as medical devices need evidence for safety and effectiveness, and hospitals still have to validate systems locally before using them in care. (fda.gov) So the release puts MedGemma 1.5 in a familiar place for health artificial intelligence in 2026: more open than many rivals, broader than a single imaging model, and still one step removed from routine bedside use. (youtube.com) (fda.gov)